Small sample reservoir hierarchical evaluation method and system based on fuzzy prior knowledge guidance

Abstract

The application relates to the technical field of oil and gas field development engineering and artificial intelligence, and discloses a small sample reservoir grading evaluation method and system based on fuzzy prior knowledge guidance, wherein the method comprises the following steps: firstly, using an entropy weight method to extract feature objective weights from data, and introducing an adaptive correction mechanism based on distribution entropy to optimize robustness; converting the optimized weights into learnable feature gate vectors, and designing a total loss function containing a KL divergence prior regularization term to deeply embed the prior knowledge into the optimization process of the neural network in an endogenous constraint manner. In this way, the small sample generalization ability is significantly improved, and the consistency of the prior framework in the training and prediction stages is realized through the learnable gating mechanism, so that the model decision path is aligned with the physical cognition, and a new paradigm of high-reliability and explainable reservoir intelligent evaluation is constructed.

Description

Technical Field

[0001] This application relates to the interdisciplinary fields of oil and gas field development engineering and artificial intelligence, such as a method and system for small-sample reservoir classification and evaluation based on fuzzy prior knowledge. Background Technology

[0002] The productivity classification and evaluation of unconventional reservoirs such as dolomite is a key basis for efficient oil and gas reservoir development and investment decisions. Currently, this field mainly faces the following four specific and interrelated technical bottlenecks:

[0003] 1. Poor overfitting and generalization under small sample sizes: Core analysis and well testing are costly, resulting in a very limited number of accurately labeled samples available for modeling (usually less than 200). Under this constraint, if a highly complex deep learning model is directly applied, the model will over-memorize noise in the training data rather than learn general patterns, leading to a sharp decline in performance on unknown data (test set or new well), i.e., generalization failure.

[0004] 2. The "Black Box" Decision-Making and Trust Crisis of Data-Driven Models: While models such as neural networks possess powerful nonlinear fitting capabilities, their internal decision-making logic is difficult to trace. In high-risk, high-cost engineering decisions such as oil and gas exploration and development, "black box" predictions lacking physical or geological support are unlikely to be adopted by domain experts, severely hindering the practical application of AI technology.

[0005] 3. Superficial integration of traditional physical models and data-driven models: Some existing studies attempt to combine the two, but most remain at the stage of process sequencing (such as using a physical model to select features first, and then inputting them into the AI ​​model) or post-hoc comparison of results. This "physics is physics, data is data" approach fails to deeply embed domain knowledge as a guiding and constraining prior information into the learning mechanism of the AI ​​model training process, and cannot achieve synergy and enhancement at the principle level;

[0006] 4. Inefficient feature utilization: In small sample scenarios, if the importance of features is not distinguished prior to their importance, the model may be interfered with by a large number of redundant or secondary features, exacerbating the risk of overfitting. Traditional feature selection methods are often independent of subsequent models and lack linkage optimization with model training.

[0007] Therefore, there is an urgent need for a new method for intelligent reservoir evaluation that can deeply embed interpretable domain knowledge into the AI ​​learning kernel, thereby simultaneously solving small sample overfitting, improving model credibility, and achieving true fusion.

[0008] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0009] To provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. This summary is not a general commentary, nor is it intended to identify key / important components or describe the scope of protection of these embodiments, but rather serves as a prelude to the detailed description that follows.

[0010] This disclosure provides a method and system for small-sample reservoir classification and evaluation based on fuzzy prior knowledge, so as to simultaneously obtain high generalization performance and high decision interpretability under small sample conditions.

[0011] In some embodiments, the small-sample reservoir grading evaluation method guided by fuzzy prior knowledge includes: acquiring multi-source features of the target reservoir and corresponding expert grading labels to construct a standardized training set matrix; calculating the objective weight vector of each feature using the entropy weight method in the fuzzy comprehensive evaluation method; calculating the distribution entropy of the objective weight vector and evaluating the degree of concentration and dispersion of the distribution entropy; adaptively correcting the objective weight vector according to the degree of concentration and dispersion; and normalizing the corrected objective weights to obtain a robust feature weight vector. The adaptive correction of the objective weight vector based on the degree of concentration and dispersion includes: if the distribution entropy representation is too dispersed, the objective weight vector is corrected using a preset geological feature prior importance vector, which is obtained based on expert experience scoring; if the distribution entropy representation is too concentrated, the objective weight vector is corrected using feature-label mutual information vector; if the distribution entropy representation is appropriately distributed, no correction is needed; the training set matrix is ​​scaled to generate a prior guiding feature matrix; the prior guiding feature matrix is ​​used as the model input to construct and train a neural network model; the total loss function includes a prior regularization term and uses the robust feature weight vector as a constraint; the features of the new well to be tested are processed and input into the trained neural network model to obtain the grading results.

[0012] In some embodiments, the small-sample reservoir grading evaluation system guided by fuzzy prior knowledge includes: a data preprocessing and quality control module, configured to acquire multi-source features of the target reservoir and corresponding expert grading labels to construct a standardized training set matrix; a fuzzy prior knowledge quantification extraction and optimization module, configured to calculate the objective weight vector of each feature using the entropy weight method in the fuzzy comprehensive evaluation method; calculate the distribution entropy of the objective weight vector and evaluate the degree of concentration and dispersion of the distribution entropy; adaptively correct the objective weight vector according to the degree of concentration and dispersion; and normalize the corrected objective weights to obtain a robust feature weight vector. The adaptive correction of the objective weight vector based on the degree of concentration and dispersion includes: if the distribution entropy representation is too dispersed, the objective weight vector is corrected using a preset geological feature prior importance vector, which is obtained based on expert experience scoring; if the distribution entropy representation is too concentrated, the objective weight vector is corrected using feature-label mutual information vector; if the distribution entropy representation is appropriately distributed, no correction is required; the prior-guided deep model endogenous training module is configured to scale the training set matrix to generate a prior-guided feature matrix; the prior-guided feature matrix is ​​used as the model input to construct and train the neural network model; the total loss function includes a prior regularization term and uses the robust feature weight vector as a constraint; the robust prediction and interpretability output module is configured to process the features of the new well to be tested and input them into the trained neural network model to obtain the grading results.

[0013] The method and system for small-sample reservoir classification and evaluation based on fuzzy prior knowledge provided in this disclosure can achieve the following technical effects:

[0014] 1. Fundamentally alleviates overfitting with small samples: By using prior weights to scale and regularize the feature space, a "learning guide" based on objective data patterns is provided to the model before and during training, significantly constraining the model's hypothesis space. This enables the neural network to learn key patterns more efficiently with limited samples, rather than fitting noise. Example data shows that on 107 total samples, this method improves the absolute accuracy by more than 9 percentage points and the relative accuracy by more than 16% compared to the same network structure without prior input.

[0015] 2. A "white-box" AI decision-making path was constructed: the model's input features have been preprocessed with weights that have clear physical meaning (such as higher weights for dynamic reserves and permeability). Therefore, any decision made by the neural network can be traced back to these weighted features, establishing a strong correlation between complex nonlinear decisions and the expert's first-order linear cognition (feature importance). This greatly enhances the credibility and acceptability of the results, providing a practical example for the application of "explainable artificial intelligence (XAI)" in the engineering field.

[0016] 3. Achieved endogenous deep fusion of "mechanism-data": This method does not simply concatenate two independent models, but rather integrates the "mechanism cognitive output" (robust feature weight vector) of fuzzy evaluation. This transforms into key constraints affecting the "training dynamics" of deep learning. This fusion, achieved through learnable gating and regularized loss, occurs within the internal loop of model training, enabling knowledge to guide the entire learning process—a true synergistic innovation where "1+1>2."

[0017] 4. A transferable and general technical framework has been formed: The fusion paradigm of "objective weights as prior transformation operators of feature space" revealed by this method has high versatility. This framework is not limited to entropy weight method and MLP, but can be seamlessly extended to other evaluation methods such as AHP and CRITIC, as well as complex network architectures such as CNN and Transformer, providing an innovative technical blueprint for solving the problem of small-sample, high-reliability AI modeling in various industries.

[0018] The above general description and the description below are exemplary and illustrative only and are not intended to limit this application. Attached Figure Description

[0019] One or more embodiments are illustrated by way of example with reference to the accompanying drawings. These illustrations and drawings do not constitute a limitation on the embodiments. Elements having the same reference numerals in the drawings are shown as similar elements. The drawings are not to be scaled. And wherein:

[0020] Figure 1 This is a schematic diagram of a small-sample reservoir classification and evaluation method based on fuzzy prior knowledge provided in an embodiment of this disclosure;

[0021] Figure 2 This is a schematic diagram of a method provided in this disclosure for calculating the objective weight vector of each feature by using the entropy weight method in the fuzzy comprehensive evaluation method to calculate the training set matrix.

[0022] Figure 3 This is a schematic diagram comparing the accuracy and overfitting of different methods provided in the embodiments of this disclosure;

[0023] Figure 4 This is a schematic diagram illustrating the comparison of feature importance weights provided in the embodiments of this disclosure;

[0024] Figure 5 This is a convergence curve of the loss function and KL divergence during the training process provided in this embodiment of the disclosure;

[0025] Figure 6 This is a schematic diagram comparing the performance of different evaluation methods and network architectures provided in the embodiments of this disclosure;

[0026] Figure 7 This is a schematic diagram of a method provided in this embodiment of the present disclosure for processing the features of a new well to be tested, inputting them into a trained neural network model, and obtaining a graded result.

[0027] Figure 8 This is a schematic diagram of a small-sample reservoir grading and evaluation system based on fuzzy prior knowledge provided in an embodiment of this disclosure.

[0028] Figure label:

[0029] 80. Small-sample reservoir classification and evaluation system guided by fuzzy prior knowledge; 81. Data preprocessing and quality control module; 82. Fuzzy prior knowledge quantification extraction and optimization module; 83. Prior-guided deep model endogenous training module; 84. Robust prediction and interpretable output module. Detailed Implementation

[0030] To provide a more detailed understanding of the features and technical content of the embodiments of this disclosure, the implementation of the embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. The accompanying drawings are for illustrative purposes only and are not intended to limit the embodiments of this disclosure. In the following technical description, for ease of explanation, several details are used to provide a full understanding of the disclosed embodiments. However, one or more embodiments may still be implemented without these details. In other cases, well-known structures and devices may be simplified in their depiction to simplify the drawings.

[0031] The terms "first," "second," etc., used in the specification and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this disclosure described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion.

[0032] Unless otherwise stated, the term "multiple" means two or more.

[0033] In this embodiment of the disclosure, the character " / " indicates that the objects before and after it are in an "or" relationship. For example, A / B means: A or B.

[0034] The term "and / or" describes an association between objects, indicating that three relationships can exist. For example, A and / or B means: A or B, or A and B.

[0035] The term "correspondence" can refer to an association or binding relationship. The correspondence between A and B means that there is an association or binding relationship between A and B.

[0036] Combination Figure 1 As shown, this disclosure provides a small-sample reservoir classification and evaluation method guided by fuzzy prior knowledge, including:

[0037] S10: Obtain the multi-source features of the target reservoir and the corresponding expert classification labels to construct a standardized training set matrix;

[0038] S20. The entropy weight method in the fuzzy comprehensive evaluation method is used to calculate the objective weight vector of each feature in the training set matrix. The distribution entropy of the objective weight vector is calculated, and the degree of concentration and dispersion of the distribution entropy is evaluated. Based on the degree of concentration and dispersion, the objective weight vector is adaptively corrected to obtain the robust feature weight vector. ;

[0039] S30, scale the training set matrix to generate a priori guided feature matrix; use the priori guided feature matrix as model input to construct and train a neural network model, where the total loss function includes a priori regularization term and uses the robust feature weight vector as a constraint.

[0040] S40 processes the features of the new well to be tested and inputs them into the trained neural network model to obtain the grading results.

[0041] Step 1: Enhanced preprocessing for small sample geological datasets

[0042] Multi-source characteristics of the target reservoir were collected, encompassing both static physical properties and dynamic production characteristics. Static physical properties included porosity, permeability, and formation coefficients. Dynamic production characteristics included average daily gas production and dynamic reserves over the past three years. Simultaneously, expert-level labels corresponding to these characteristics were acquired. These labels were independently assigned production capacity levels (e.g., I, II, III, IV) by multiple senior geological engineers. Consistent results were then used to obtain multiple high-confidence samples, forming a small sample dataset.

[0043] The small sample size means that each sample is extremely valuable, making data quality control crucial. The following preprocessing method is used to preprocess the small samples to obtain a standardized training set matrix. Preprocessing includes:

[0044] (1) Missing value handling: For random missing values, the feature mean is used for filling; for systematic missing values ​​related to geological strata, spatial interpolation is performed by combining data from adjacent wells. This combined strategy takes into account both efficiency and geological rationality;

[0045] (2) Outlier Handling: The simple 3σ criterion is abandoned (because the distribution assumption is unreliable with small samples), and a robust truncation method based on the interquartile range (IQR) is adopted. For each feature, its upper quartile is calculated (…). ) and lower quartiles ( ), will exceed Values ​​within this range are considered geological anomalies or measurement errors and replaced with the corresponding boundary values. This method is less sensitive to extreme values ​​in small samples and is more robust.

[0046] (3) Data standardization: Z-score standardization is adopted. The key is the mean. and standard deviation The calculations are performed only on the training set and are fixed for the test set and future new samples. This simulates real-world application scenarios, prevents data leakage, and is a prerequisite for evaluating small-sample generalization ability.

[0047] In this set, a portion of the small sample constitutes the training set, and the other portion constitutes the test set. For example, if 107 high-confidence samples are ultimately obtained, then 55 samples constitute the training set and 52 samples constitute the test set.

[0048] Step 2: Quantitative Extraction and Robustness Optimization of Interpretable Prior Knowledge Based on Entropy Weight Method

[0049] The core of this step is to objectively and quantitatively extract the importance of features implicit in the data, and optimize their distribution through an adaptive correction mechanism to form robust prior knowledge that can be embedded in subsequent models. Specifically, this includes:

[0050] S21, Primary Objective Weight Calculation: The entropy weight method in fuzzy comprehensive evaluation is used to calculate the objective weight vector of each feature from the standardized training set matrix obtained in S10. Combined with... Figure 2 As shown, it specifically includes:

[0051] S211, For the standardized training set matrix, calculate its i-th... Information entropy of each feature :

[0052] ,

[0053] Where n is the number of samples, , Among them, for Perform a tiny translation to make all of them positive;

[0054] S212, Calculate the coefficient of difference : . The larger the value, the greater the difference in this feature among different samples, and the greater its potential contribution to classification.

[0055] S213, regarding the coefficient of difference Normalization is performed to obtain the primary weights. This leads to the formation of an objective weight vector. .

[0056] Table 1. Examples of key feature weights calculated using the entropy weight method.

[0057]

[0058] In this way, the initial objective quantification of feature importance is achieved through the entropy weight method of S21.

[0059] S22, Weight Distribution Quality Control: Calculate the distribution entropy of the objective weight vector and assess the degree of concentration or dispersion of the distribution entropy. The degree of concentration or dispersion includes three states: overly concentrated, overly dispersed, and appropriately distributed. Specifically:

[0060] First, calculate the distribution entropy of the objective weight vector:

[0061] ;

[0062] in, For objective weight vectors, m It is the characteristic number.

[0063] Set the first distribution entropy threshold Second distribution entropy threshold , > .pass , The distribution entropy can be divided into concentrated and dispersed types. Optional. The value is 0.7. The value is 1.2.

[0064] Then and , Comparison:

[0065] if If so, it is determined that the distribution entropy is too concentrated;

[0066] if If so, it is determined that the distribution entropy is too dispersed;

[0067] if If the distribution entropy is suitable, then the distribution entropy distribution is determined to be appropriate.

[0068] In this way, by calculating the distribution entropy and comparing it with the threshold, a quantitative assessment and anomaly identification of the objective weight distribution state are achieved, providing a clear decision basis for subsequent adaptive correction.

[0069] S23, Adaptive weight correction: Adaptively correct the objective weight vector based on the evaluation results.

[0070] S231, Smoothing correction for "over-concentration": Introducing a pre-defined prior importance vector of geological features. It is based on expert experience scoring and normalized to The objective weight vector is corrected using a convex combination:

[0071] ;

[0072] in, This represents the corrected objective weights. Represents the coefficient.

[0073] Thus, when the weights are excessively concentrated, and Weighted smoothing is applied to enhance the smoothing effect of geological priors and prevent the model from ignoring minor but important features.

[0074] S232, Focus correction for "overly scattered" features: Calculate the mutual information between each feature and the label. This forms a vector I, which is then normalized to... The objective weight vector is corrected using a convex combination:

[0075] ;

[0076] in, This represents the corrected objective weights. Represents the coefficient.

[0077] Thus, when the weights are too dispersed, and We will implement weighted focusing, strengthen the weight of features that are strongly correlated with the target, and improve the pertinence of our guidance.

[0078] S233, regarding "appropriate distribution": no correction is needed, let .

[0079] S234, to Normalization is performed to obtain a robust feature weight vector for subsequent guidance.

[0080] .

[0081] In this way, by correcting and normalizing the objective weight vector through S23, the objective quantitative extraction of interpretable prior knowledge can be achieved.

[0082] In summary, S20 employs the entropy weight method in fuzzy comprehensive evaluation to calculate the primary objective weight vector on the preprocessed training data; it then introduces a weight distribution entropy evaluation and adaptive correction mechanism to generate a robust feature weight vector. , As physical prior knowledge that is extracted from and optimized from the data itself and can be mathematically interpreted.

[0083] Step 3: Training guided by the endogenous prior knowledge based on learnable gating and regularization constraints

[0084] This step is crucial for achieving "deep organic fusion," and it involves transforming the robust feature weight vector... It is deeply integrated into the optimization process of neural networks in a learnable, interpretable, and stable manner.

[0085] S31, Construct learnable feature gating vectors:

[0086] Initialize the learnable parameter vector .

[0087] Feature gating vectors are constructed using the Sigmoid function. :

[0088] ;

[0089] To achieve initial boot, Initialize to:

[0090] ;

[0091] Make .

[0092] S32, Design a priori-guided regularized total loss function :

[0093] ;

[0094] in, Cross-entropy is used for classification loss.

[0095] As the prior regularization term, the best option is... Divergence constraint:

[0096] ;

[0097] Alternatively, a correlation-sparseness composite constraint is preferred:

[0098] ;

[0099] The regularization coefficient is . is the sparsity coefficient. Optionally, The value is 0.1. The value is 0.01.

[0100] S33, Forward Propagation and Feature Transformation: For the standardized training set matrix obtained in S10 X Using the current feature gating vector Scaling:

[0101] ;

[0102] The prior guiding feature matrix serves as the input to the neural network model; This indicates element-wise multiplication.

[0103] S34, Backpropagation and Gradient Adjustment Mechanism (Endogenous Guidance Core): The gradient of the total loss function with respect to V is:

[0104] ;

[0105] Drive Adjustments should be made to reduce classification errors.

[0106] Drive Preservation and a priori Consistent.

[0107] The interplay between these two gradients enables the prior to provide dynamic and gentle guidance for the optimization trajectory.

[0108] S35, Training Stability Guarantee: Due to Its derivative is bounded, fundamentally avoiding gradient explosion / vanishing. Monitoring during training. If the value continues to rise, it may trigger an early shutdown.

[0109] S36, Network Architecture and Training: Construct a CBAM-MLP network. MLP was chosen because its fully connected nature makes it easy to demonstrate the influence of feature weights; CBAM (Convolutional Block Attention Module) is introduced to enhance the model's ability to further focus on important information in the weighted features, forming a dual guidance mechanism of "global prior weights + local adaptive attention". The specific structure is as follows:

[0110] Input (13-dimensional) → FC1 (128) → LayerNorm → ReLU → Dropout (0.3) → FC2 (64) → LayerNorm → ReLU → Dropout (0.3) → FC3 (32) → LayerNorm → ReLU → Dropout (0.3) → CBAM module → Output layer (4, Softmax).

[0111] Dropout and LayerNorm are set up to further improve stability during training with few samples.

[0112] Optimizer: L-BFGS is selected. In mini-batch or full-batch settings, L-BFGS, due to its accurate second-order approximation capability, often converges faster and more stably on small datasets than first-order optimizers (such as SGD and Adam).

[0113] Training regime: 100 epochs, with early stopping (patience=10) used to monitor validation set loss.

[0114] S37, Training complete: Optimal feature gating obtained. and network parameters. It contains prior knowledge for prediction.

[0115] In this way, by using the robust feature weight vector obtained from S20 as a constraint through S30, and by designing a total loss function that includes a priori regularization term, a learnable feature gating mechanism is constructed to realize the endogeneity and dynamic guidance of prior knowledge on the direction of model optimization.

[0116] Step 4: Systematic Performance Verification and Interpretability Analysis

[0117] To comprehensively evaluate the evaluation method of this embodiment, rigorous comparative experiments were conducted.

[0118] 1. Comparative experimental design:

[0119] Baseline A (pure mechanism): linear weighted scoring and classification using only the entropy weight method results.

[0120] Baseline B (purely data-driven): the same CBAM-MLP, inputting raw standardized features (without prior weighting and gating mechanisms).

[0121] The method in this embodiment is to execute S10 to S30 completely.

[0122] 2. Quantitative performance results:

[0123] Table 2 Performance Comparison of Test Set (52 samples)

[0124]

[0125] See Figure 3 And Table 2, Figure 3 The middle horizontal axis represents the three evaluation methods (baseline A fuzzy evaluation, baseline B original neural network, and our method), the left vertical axis represents the test accuracy (bar chart), and the right vertical axis represents the training-test accuracy difference (line chart). From Figure 3 As shown in the figure, the test accuracy of this method reaches 67.31%, an absolute improvement of 9.62 percentage points compared to baseline B. Meanwhile, the training-test accuracy difference is only 5 percentage points, far less than the 15 percentage points of baseline B. This figure directly proves technical effect 1: this method can significantly alleviate overfitting on small samples, greatly narrowing the generalization gap while improving accuracy.

[0126] 3. Empirical Explanation of Interpretability: Test samples were selected, and the decision-making basis for classifying them as Type I wells according to the model in this embodiment was analyzed. (See also...) Figure 4 , Figure 4 The horizontal axis represents the normalized weights, and the vertical axis represents the top ten reservoir features ranked by expert prior knowledge. The two sets of horizontal bars represent the importance of expert prior knowledge and the gating vector after convergence of this method, respectively. The figure shows that the weights of the top three features (dynamic reserves, permeability 2, and average daily gas production in the first three years) are highly consistent between the two, and the overall ranking and numerical distribution errors are extremely small. This figure directly proves technical effect 2: This method embeds expert cognition into the input layer in the form of endogenous gating, aligning the neural network decision path with physical priors, and realizing "white-box" interpretable AI.

[0127] 4. Deep Fusion Verification: See [link / reference] Figure 5 , Figure 5 The middle horizontal axis represents the number of training epochs, and the left vertical axis represents the classification loss. Total loss The right vertical axis represents the KL divergence. The legend distinguishes the three curves by different line types. The figure shows that the KL divergence monotonically decreases from its initial high level and approaches zero, while the classification loss decreases and converges synchronously. This figure directly proves technical effect 3: This method deeply embeds domain knowledge into the training dynamics through prior regularization terms, achieving an endogenous fusion of mechanism cognition and data-driven approaches, rather than a simple concatenation.

[0128] 5. Transferable general verification: See Figure 6 , Figure 6 The horizontal axis represents the five method combinations (entropy weight method + MLP, AHP + MLP, CRITIC + MLP, entropy weight method + CNN, and entropy weight method + Transformer), and the vertical axis represents the test accuracy. Each pair of bars represents the proposed method framework and the original baseline without prior knowledge, respectively. The figure shows that the accuracy of the proposed method framework is consistently higher than 65% in all combinations, representing an absolute improvement of 8 to 10 percentage points compared to the baseline. This figure directly demonstrates technical effect 4: the proposed paradigm of "objective weights as prior transformation operators for feature space" has high versatility and can be seamlessly transferred to various evaluation methods and complex network architectures, providing a reusable technical blueprint for solving the problem of high-reliability modeling with small samples.

[0129] Step 5: Robust and interpretable prediction process for new samples.

[0130] Combination Figure 7 As shown, the application phase process is clear and ensures consistency between training and application:

[0131] S41, Data homogenization: The mean is saved using the training set matrix that has been standardized in S10. and standard deviation Standardize the logging features. This is crucial for ensuring model interpretability, guaranteeing that the features used for prediction have the same physical scale as those used during training.

[0132] S42, Prior Consistency Transformation: Using the optimal feature gating vector fixed after training. Perform feature weighting: . This represents the standardized feature vector of the new well to be tested.

[0133] S43, Model Prediction and Report Generation: [This section appears to be incomplete and requires further context.] The data is input into the trained neural network model to obtain the ranking results. The system also outputs the original values, weighted values, and their contribution ranking to the prediction results for each feature of the sample, forming a reliable evaluation report.

[0134] In this way, through S40, for the sample to be predicted, feature transformation is performed using the same preprocessing parameters as in the training phase and a fixed feature gating vector after training convergence. This transformed feature is then input into the trained neural network model to obtain the final classification result. This process ensures that the training and prediction phases follow the exact same prior knowledge framework, guaranteeing the consistency of robustness and interpretability in model application.

[0135] The small-sample reservoir classification and evaluation method based on fuzzy prior knowledge provided in this disclosure has the following beneficial effects:

[0136] 1. Fundamentally alleviates overfitting with small samples: By using prior weights to scale and regularize the feature space, a "learning guide" based on objective data patterns is provided to the model before and during training, significantly constraining the model's hypothesis space. This enables the neural network to learn key patterns more efficiently with limited samples, rather than fitting noise. Example data shows that on 107 total samples, this method improves the absolute accuracy by more than 9 percentage points and the relative accuracy by more than 16% compared to the same network structure without prior input.

[0137] 2. A "white-box" AI decision-making path was constructed: the model's input features have been preprocessed with weights that have clear physical meaning (such as higher weights for dynamic reserves and permeability). Therefore, any decision made by the neural network can be traced back to these weighted features, establishing a strong correlation between complex nonlinear decisions and the expert's first-order linear cognition (feature importance). This greatly enhances the credibility and acceptability of the results, providing a practical example for the application of "explainable artificial intelligence (XAI)" in the engineering field.

[0138] 3. Achieved endogenous deep fusion of "mechanism-data": This method does not simply concatenate two independent models, but rather integrates the "mechanism cognitive output" (robust feature weight vector) of fuzzy evaluation. This transforms into key constraints affecting the "training dynamics" of deep learning. This fusion, achieved through learnable gating and regularized loss, occurs within the internal loop of model training, enabling knowledge to guide the entire learning process—a true synergistic innovation where "1+1>2."

[0139] 4. A transferable and general technical framework has been formed: The fusion paradigm of "objective weights as prior transformation operators of feature space" revealed by this method has high versatility. This framework is not limited to entropy weight method and MLP, but can be seamlessly extended to other evaluation methods such as AHP and CRITIC, as well as complex network architectures such as CNN and Transformer, providing an innovative technical blueprint for solving the problem of small-sample, high-reliability AI modeling in various industries.

[0140] Combination Figure 8 As shown in the figure, this disclosure discloses a small-sample reservoir grading and evaluation system 80 based on fuzzy prior knowledge, including: a data preprocessing and quality control module 81, a fuzzy prior knowledge quantification extraction and optimization module 82, a prior-guided deep model endogenous training module 83, and a robust prediction and interpretability output module 84. The data preprocessing and quality control module 81 is configured to acquire multi-source features of the target reservoir and corresponding expert grading labels to construct a standardized training set matrix. The fuzzy prior knowledge quantification extraction and optimization module 82 is configured to use the entropy weight method in the fuzzy comprehensive evaluation method to calculate the objective weight vector of each feature; calculate the distribution entropy of the objective weight vector and evaluate the degree of concentration and dispersion of the distribution entropy; adaptively correct the objective weight vector according to the degree of concentration and dispersion; and normalize the corrected objective weights to obtain a robust feature weight vector. The adaptive correction of the objective weight vector based on the degree of concentration and dispersion includes: if the distribution entropy representation is too dispersed, the objective weight vector is corrected using a preset geological feature prior importance vector, which is obtained based on expert experience scoring; if the distribution entropy representation is too concentrated, the objective weight vector is corrected using a feature-label mutual information vector; if the distribution entropy representation is appropriately distributed, no correction is needed. The prior-guided deep model endogenous training module 83 is configured to scale the training set matrix to generate a prior-guided feature matrix; the prior-guided feature matrix is ​​used as model input to construct and train the neural network model; the total loss function includes a prior regularization term and uses the robust feature weight vector as a constraint. The robust prediction and interpretability output module 84 is configured to process the features of the new well to be tested, input them into the trained neural network model, and obtain the grading results.

[0141] The specific implementation process of this system can be found in the description of the above method embodiments, and will not be repeated here.

[0142] The technical solutions of this disclosure can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes one or more instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of this disclosure. The aforementioned storage medium can be a non-transitory storage medium, such as a USB flash drive, external hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, etc., and other media capable of storing program code.

[0143] The foregoing description and accompanying drawings fully illustrate embodiments of this disclosure to enable those skilled in the art to practice them. Other embodiments may include structural, logical, electrical, procedural, and other changes. The embodiments represent only possible variations. Individual components and functions are optional unless explicitly required, and the order of operation may vary. Parts and features of some embodiments may be included or replace parts and features of other embodiments. Without further limitations, an element defined by the phrase "comprising a…" does not exclude the presence of additional identical elements in the process, method, or apparatus that includes the element. Throughout this document, each embodiment may focus on its differences from other embodiments, and similar or identical parts between embodiments may be referred to mutually. For methods, products, etc., disclosed in the embodiments, if they correspond to the method section disclosed in the embodiments, then the relevant parts may be referred to the description of the method section.

[0144] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the embodiments of this disclosure. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0145] The methods and products (including but not limited to devices and equipment) disclosed in the embodiments herein can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of units may be merely a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces, and the indirect coupling or communication connection of devices or units may be electrical, mechanical, or other forms. Units described as separate components may or may not be physically separate, and components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units may be selected to implement this embodiment according to actual needs. Furthermore, the functional units in the embodiments of this disclosure may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0146] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. In the descriptions corresponding to the flowcharts and block diagrams in the accompanying drawings, the operations or steps corresponding to different blocks may also occur in a different order than disclosed in the description, and sometimes there is no specific order between different operations or steps. For example, two consecutive operations or steps may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. Each block in a block diagram and / or flowchart, and combinations of blocks in a block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

Claims

1. A small-sample reservoir classification and evaluation method based on fuzzy prior knowledge, characterized in that, include: Obtain multi-source features of the target reservoir and corresponding expert classification labels to construct a standardized training set matrix; The entropy weight method in fuzzy comprehensive evaluation is used to calculate the objective weight vector of each feature from the training set matrix. The distribution entropy of the objective weight vector is calculated, and the degree of concentration or dispersion of the distribution entropy is evaluated. Based on the degree of concentration or dispersion, the objective weight vector is adaptively corrected, and the corrected objective weights are normalized to obtain the robust feature weight vector. The adaptive correction of the objective weight vector based on the degree of concentration and dispersion includes: if the distribution entropy representation is too dispersed, then the objective weight vector is corrected using a preset geological feature prior importance vector, wherein the preset geological feature prior importance vector is obtained based on expert experience scoring; if the distribution entropy representation is too concentrated, then the objective weight vector is corrected using a feature-label mutual information vector; if the distribution entropy representation is appropriately distributed, then no correction is required. The training set matrix is ​​scaled to generate a prior guided feature matrix; the prior guided feature matrix is ​​used as the model input to construct and train a neural network model; wherein the total loss function includes a prior regularization term and the robust feature weight vector is used as a constraint condition. The features of the new well to be tested are processed and input into the trained neural network model to obtain the classification results; The step of correcting the objective weight vector using a preset geological feature prior importance vector includes: ， in, This represents the corrected objective weights; , representing the pre-defined a priori importance vector of geological features, which is formed after normalization. ; , represents the coefficient. This represents the first distribution entropy threshold; The step of correcting the objective weight vector using feature-label mutual information vectors includes: ， in, This represents the corrected objective weights; This represents the feature-label mutual information vector; Represents the coefficient. This represents the second distribution entropy threshold; The scaling of the training set matrix to generate the prior guided feature matrix includes: Initialize the learning parameter vector Through the Sigmoid function Constructing feature gating vectors : ； Will Initialize to: ； Make ; The training set matrix is ​​used with the feature gating vector G. Scaling: ， The prior guiding feature matrix is ​​used as the input to the neural network model. This indicates element-wise multiplication.

2. The small-sample reservoir classification and evaluation method based on fuzzy prior knowledge as described in claim 1, characterized in that, The entropy weight method in the fuzzy comprehensive evaluation method is used to calculate the objective weight vector of each feature on the training set matrix, including: For the training set, calculate its i.e. Information entropy of each feature : ， Where n is the number of samples, , Among them, for Perform a tiny translation to make all of them positive; Calculate the difference coefficient based on the information entropy. : ; The difference coefficients are normalized to obtain the primary weights. To form the objective weight vector .

3. The small-sample reservoir classification and evaluation method based on fuzzy prior knowledge as described in claim 1, characterized in that, The calculation of the distribution entropy of the objective weight vector and the evaluation of the concentration and dispersion of the distribution entropy include: ， in, Let the objective weight vector be... For primary weights, m It is the characteristic number.

4. The small-sample reservoir classification and evaluation method based on fuzzy prior knowledge as described in claim 1, characterized in that, The assessment of the concentration and dispersion of the distribution entropy includes: If the distribution entropy is less than the first distribution entropy threshold, then the distribution entropy is determined to be too concentrated. If the distribution entropy is greater than the second distribution entropy threshold, then the distribution entropy is determined to be too dispersed; If the distribution entropy is greater than or equal to the first distribution entropy threshold and less than or equal to the second distribution entropy threshold, then the distribution entropy distribution is determined to be suitable. Wherein, the first distribution entropy threshold is less than the second distribution entropy threshold.

5. The small-sample reservoir classification and evaluation method based on fuzzy prior knowledge as described in claim 1, characterized in that, The construction and training of the neural network model includes: The total loss function is: ， in, Cross-entropy is used for classification loss; The prior regularization term is represented as: ;or, ， in, The regularization coefficient is . The sparsity coefficient; The total loss function is... V The gradient is: ， Drive Adjust towards reducing classification errors. Drive Preservation and a priori Consistency is achieved, thus obtaining the optimal feature gating vector. .

6. The method for small-sample reservoir classification and evaluation based on fuzzy prior knowledge according to any one of claims 1 to 5, characterized in that, The process of processing the features of the new well to be tested and inputting them into the trained neural network model includes: The features of the new well to be tested are standardized using the same mean and standard deviation as the training set matrix. Using the fixed optimal feature gating vector after training Perform feature weighting: ， in, This indicates element-wise multiplication; This represents the standardized feature vector of the new well to be tested; Will The input is fed into the trained neural network model.

7. A small-sample reservoir classification and evaluation system guided by fuzzy prior knowledge, characterized in that, include: The data preprocessing and quality control module is configured to acquire multi-source features of the target reservoir and corresponding expert classification labels in order to construct a standardized training set matrix. The fuzzy prior knowledge quantification extraction and optimization module is configured to use the entropy weight method in the fuzzy comprehensive evaluation method to calculate the objective weight vector of each feature on the training set matrix; calculate the distribution entropy of the objective weight vector and evaluate the concentration and dispersion of the distribution entropy; adaptively correct the objective weight vector according to the concentration and dispersion; and normalize the corrected objective weights to obtain the robust feature weight vector. The adaptive correction of the objective weight vector based on the degree of concentration and dispersion includes: if the distribution entropy representation is too dispersed, then the objective weight vector is corrected using a preset geological feature prior importance vector, wherein the preset geological feature prior importance vector is obtained based on expert experience scoring; if the distribution entropy representation is too concentrated, then the objective weight vector is corrected using a feature-label mutual information vector; if the distribution entropy representation is appropriately distributed, then no correction is required. The prior-guided deep model intrinsic training module is configured to scale the training set matrix to generate a prior-guided feature matrix; use the prior-guided feature matrix as model input to construct and train a neural network model; wherein the total loss function includes a prior regularization term and uses the robust feature weight vector as a constraint condition. The robust prediction and interpretable output module is configured to process the features of the new well to be tested, input them into the trained neural network model, and obtain the hierarchical results. The step of correcting the objective weight vector using a preset geological feature prior importance vector includes: ， in, This represents the corrected objective weights; , representing the pre-defined a priori importance vector of geological features, which is formed after normalization. ; , represents the coefficient. This represents the first distribution entropy threshold; The step of correcting the objective weight vector using feature-label mutual information vectors includes: ， in, This represents the corrected objective weights; This represents the feature-label mutual information vector; Represents the coefficient. This represents the second distribution entropy threshold; The scaling of the training set matrix to generate the prior guided feature matrix includes: Initialize the learning parameter vector Through the Sigmoid function Constructing feature gating vectors : ； Will Initialize to: ； Make ; The training set matrix is ​​used with the feature gating vector G. Scaling: ， The prior guiding feature matrix is ​​used as the input to the neural network model. This indicates element-wise multiplication.

Images